Single Domain Generalization for Crowd Counting

This is an official repository for our CVPR2024 work, "Single Domain Generalization for Crowd Counting". You can read our paper here.

Requirements

• Python 3.10.12
• PyTorch 2.0.1
• Torchvision 0.15.2
• Others specified in requirements.txt

Data Preparation

1. Download ShanghaiTech and UCF-QNRF datasets from official sites and unzip them.
2. Run the following commands to preprocess the datasets:

   python utils/preprocess_data.py --dataset sta --origin-dir [path_to_ShanghaiTech]/part_A --data-dir data/sta
   python utils/preprocess_data.py --dataset stb --origin-dir [path_to_ShanghaiTech]/part_B --data-dir data/stb
   python utils/preprocess_data.py --dataset qnrf --origin-dir [path_to_UCF-QNRF] --data-dir data/qnrf
   

3. Run the following commands to generate GT density maps:

   python utils/dmap_gen.py --path data/sta
   python utils/dmap_gen.py --path data/stb
   python utils/dmap_gen.py --path data/qnrf
   

Training

Run the following command:

python main.py --task train --config configs/sta_train.yml

You may edit the .yml config file as you like.

Testing

Run the following commands after you specify the path to the model weight in the config file:

python main.py --task test --config configs/sta_test_stb.yml
python main.py --task test --config configs/sta_test_qnrf.yml

Inference

Run the following command:

python inference.py --img_path [path_to_img_file_or_directory] --model_path [path_to_model_weight] --save_path output.txt --vis_dir vis

Pretrained Weights

Original Paper (Non-Deterministic)

We provide pretrained weights that produce the results in our original paper in the table below. Note that due to the use of non-deterministic torch.nn.functional.interpolate, the results are not reproducible by training from scratch.

Source	Performance	Weights
A	B: 11.4MAE, 19.7MSE Q: 115.7MAE, 199.8MSE	OneDrive Google Drive
B	A: 99.6MAE, 182.9MSE Q: 165.6MAE, 290.4MSE	OneDrive Google Drive
Q	A: 65.5MAE, 110.1MSE B: 12.3MAE, 24.1MSE	OneDrive Google Drive

New Deterministic Version

To make the results fully reproducible, we implement a new version of MPCount which does not use torch.nn.functional.interpolate. We will release more results and weights soon.

Source	Performance	Weights
A	B: 11.2MAE, 20.0MSE Q: 112.8MAE, 193.8MSE	OneDrive Google Drive

Citation

If you find this work helpful in your research, please cite the following:

@inproceedings{pengMPCount2024,
  title = {Single Domain Generalization for Crowd Counting},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024)},
  author = {Peng, Zhuoxuan and Chan, S.-H. Gary},
  year = {2024}
}